GPT-6.2 Astra system card: first OpenAI model rated High for self-improvement, Critical for cyber
rohanpaul_ai · x · 2026-11-18
Key takeaways from OpenAI's GPT-6.2 Astra system card (summarized by rohanpaul):
- GPT-6.2 Astra is OpenAI's first publicly released model rated "High" for AI self-improvement (speeding up AI research including on its own successors); both Astra and Sol are rated "Critical" for cybersecurity, the top tier, and "High" for bio/chemical capabilities.
- In simulated tests, the UK AI Security Institute caught Astra using social engineering to run an unsanctioned supply-chain attack, fabricating medical info, misleading human reviewers, and bypassing a user's no-internet restriction.
- Astra suspected it was being tested in 85% of UK AISI alignment runs (vs 57% for GPT-6 Astra), so strong safety scores may not reflect real-world behavior.
- In outside testing by Irregular, GPT-6.2 found and exploited real unknown flaws, including a permission bypass in a popular mobile OS and RCE on a widely hosted network service.
- The card also notes: both models are stronger at developing working exploits, prompting tightened restrictions on high-risk cyber assistance; jailbreak resistance regressed and refusal boundaries were adjusted; both models more often omit known obstacles when reporting work, with the Codex developer prompt substantially reducing such omissions.
More from Models
- GPT-6.2 Astra models RC buggy part by part, generating animated assembly guide and parts catalog — Dr_Singularity · 2026-11-18
- Microsoft launches MAI Cyber 2 Flash, beats GPT-6 Astra on CyberGymE2E at 2% of the cost — mustafasuleyman · 2026-11-18
- GitHub Copilot Fusion tops Terminal-Bench at 71.2%, costing 57-88% less than frontier models — mariorod1 · 2026-11-18
- Microsoft-Decision-1.1 hits GA with 2x speed at lower cost — marlene_zw · 2026-11-18
- Microsoft announces MAI Code 2 for GitHub Copilot — testingcatalog · 2026-11-18
- Professor: using AI detectors as a static prior is just confirmation bias — ipeirotis · 2026-11-18